Neural Denoisers for Low SPP DirectX Raytracing

Real-time DirectX Raytracing (DXR) offers physically accurate lighting, but hardware limits dictate that modern game engines can only cast one or two rays per pixel per frame. This low sample-per-pixel (spp) budget produces heavily undersampled, noisy images that are unplayable without post-processing. A neural denoiser bridges this gap by leveraging trained deep-learning models to reconstruct clean, temporally stable, and high-fidelity frames from noisy radiance data and engine-provided geometric guide buffers in real time.

The Challenge of Low-SPP Raytracing

DirectX Raytracing calculates effects like glossy reflections, ambient occlusion, global illumination, and soft shadows by tracing ray paths across a scene. Offline renderers cast hundreds or thousands of samples per pixel to produce clean imagery. In real-time rendering targeting 60 frames per second or higher, developers are restricted to extremely sparse sampling—often just 1 spp.

At 1 spp, the raw output from DXR is not a complete image, but a stochastic, high-frequency "salt-and-pepper" noise pattern. Standard spatial blur filters fail to clean this up effectively, as they either blur away fine geometric details or preserve noise along sharp edges.

How a Neural Denoiser Operates

Neural denoisers replace or augment traditional heuristic filters (such as bilateral, cross-bilateral, or temporal anti-aliasing filters) with specialized deep learning models, typically based on convolutional neural networks (CNNs), recurrent neural networks (RNNs), or autoencoder architectures.

Instead of evaluating pixels in isolation, the neural network analyzes multiple inputs simultaneously:

• Noisy Radiance Buffers: The raw, un-denoised color and lighting data generated by DXR passes.
• G-Buffer Auxiliary Data: High-precision geometric information, including surface normals, depth (Z-buffer), and albedo (base color).
• Motion Vectors: Pixel movement data between the current and previous frames to facilitate temporal tracking.

Trained on millions of ground-truth frames rendered with thousands of samples, the network learns the statistical patterns of how light interacts with geometry. It accurately distinguishes between high-frequency noise and intentional high-frequency scene details, such as hair, textured surfaces, and shadow borders.

Core Roles of the Neural Denoiser

1. Feature-Aware Noise Suppression

Traditional filters often mistake lighting noise for surface texture, or vice versa. Neural denoisers use auxiliary G-buffer data to infer where geometric boundaries exist. This allows the system to smooth out noise within a surface while preserving crisp edges, normal map variations, and geometry silhouettes without washing them out.

2. Temporal Accumulation and Ghosting Mitigation

Ray tracing noise fluctuates aggressively from frame to frame. Neural denoisers accumulate radiance over time, using motion vectors to reproject pixels from historical frames into the current view. While standard temporal filters struggle with rapid movements and disocclusions—often resulting in smearing, boiling artifacts, or "ghosting" trails behind moving objects—neural networks are trained to recognize when historical data is invalid, discarding stale information instantly.

3. Reconstruction of Complex Light Transport

Certain lighting effects, such as specular reflections on curved surfaces or indirect diffuse bounces, create complex, non-linear noise patterns that classical mathematics cannot easily generalize. Neural denoisers excel at pattern recognition, intelligently predicting and filling in missing lighting values to reproduce authentic surface gloss, light bleed, and contact shadows from partial data.

Enabling Real-Time Path Tracing

By shifting the workload from brute-force ray calculation to dedicated AI accelerators (such as Tensor Cores), neural denoising makes real-time ray tracing commercially viable. It reduces the computational cost of DXR by orders of magnitude, turning raw 1-spp outputs into final frames that rival the fidelity of offline path-traced renders.